Option 1:
If you are able to use a large context model like Claude Sonet etc, having one AI column would get you better results.
Option 2:
If you are pulling the signals into separate columns, you can reference only one signal for each message.
Option 3:
If you want to use the exclusion logic. I highly recommend SPICE framework. Feed that into any LLM and ask it to write the prompt. The key here are the examples.